Methods of sampling based on exhaustive and evolutionary search

Abstract

The development of mathematical software for training sampling is considered. Exhaustive and evolutionary sampling methods are developed. Criteria for selection, censoring, and pseudoclustering of instances are introduce in these methods. This makes it possible to speed up the sampling process and to ensure the compliance of the samples with the limited size. The proposed methods allow for the automatic allocation of a subset of instances with the minimal size from the original sample. The subset contains the most important instances for the model’s construction. The complexity estimates of the developed methods are defined. Experiments to determine the practical applicability of the methods are conducted. The use of the proposed estimates and identified dependences makes it possible to take into account the available computer resources during the sampling.